The current reference describes the implementation; it is not evidence that the owner endorsed every sentence. Existing prose was largely assistant-written. When a description conflicts with code, investigate the difference instead of restoring an older design automatically. Explicit owner instructions take precedence over inferred intent in documentation.
Start with:
- GID: the native logical substrate and the boundary with the temporary text bridge.
- Data and editor model: addresses, selection, completion, panes, history, and persistence.
- Grap and projections: evaluation, libraries, display composition, and host boundaries.
- Dependency-tracked computations: memo graphs, Grap observations, cancellable background jobs, and the first CAM integration.
- Puri and the editor frame: ownership, layout, events, hover, and drawing.
- Build security: sandboxed build/test commands.
- Platforms: the native and browser hosts.
Browser worker experiment records the shared-memory closure probe, the UI-thread synchronization fixes, and the browser executor/Rayon integration. Browser setup covers the host. Browser/native profiling compares the actual CAM document's computation and retained mesh drawing on both backends, including the browser worker-pool and WebGPU improvements. Browser SIMD profiling measures the smaller compiler-flag improvement and checks final color/depth outputs. Fidget bulk-loop experiment profiles the actual browser rendering workers and tests a safe slice-loop rewrite that enables vectorization; the measured patch is now applied to the vendored core. Browser orbit profiling follows the actual frame/gesture path and separates source-pane construction from background worker contention. Allocator experiments compare two alternative WASM allocator configurations against that same orbit replay.
Deferred work records the unresolved items set aside during the review. Historical notes preserve earlier models and proposals separately; they are not required reading for ordinary changes. Layout continuations describes the box/widget boundary, the frame stages, and their verification. Layout navigation describes logical-line navigation and its checkpoint measurements. The superseded neighbor callback experiment is preserved as history.
Performance checks describes the shared headless frame harness and the IoP/Fidget canaries. Tree profiling is a dated measurement report. Fidget meshing records the headless cube experiment and its quality limitations. The mesh viewport describes the subsequent in-app option. Its toolpath variant now retains dependency-tracked geometry; ordinary mesh previews still remesh each frame. Retained progressive rendering records test-only whole-view, bounded-tile, exact-program-sharing, subtree census, and persistent-expression measurements. CAM now tries mesh → final-quality implicit tiles; the standalone implicit preview retains independent refinement passes. Hybrid CAM rendering records the subsequent split: implicit model/stock color and depth combined with mesh paths and cutters. Graphics memory records the idle-footprint investigation and isolated Vello resource-retention measurements. Separate image composition records the experiment and native integration around Vello's image atlas, including clipping, upload reuse, GPU-texture handoff, and single-/multiple-preview presentation costs. Direct GPU mesh drawing now uses that same ordered compositor on native and WebGPU; the browser retains Canvas2D/CPU drawing as a fallback. GPU spills records the paused, opt-in GPU interpreter experiment, CPU comparisons, program-arena pressure, and batching tradeoffs. CAM remains on the CPU renderer. Library resolution records the later definition-storage change and its performance checks. The release checklist covers distribution concerns.
Examples lists the bundled documents, shortcuts, and the formulas behind the Fidget shapes. Toolpaths describes the first streaming CAM geometry library and its Rhino-derived combined model-and-path preview. Tool profiles describes revolved cutter sections and the remaining tool-change/chamfer work. Tool-sweep experiment records the test-only analytic bull-mill candidate and why it has not replaced the working approximation. Controls describes reusable sliders and Grap control composition used by toolpath playback. Final-encoded trees describes scoped hierarchy construction and source-linked collection shared by the CAM controls and preview.